* AAC encoder psychoacoustic model
*/
+#include "libavutil/attributes.h"
#include "avcodec.h"
#include "aactab.h"
#include "psymodel.h"
* information for single band used by 3GPP TS26.403-inspired psychoacoustic model
*/
typedef struct AacPsyBand{
- float energy; ///< band energy
- float thr; ///< energy threshold
- float thr_quiet; ///< threshold in quiet
+ float energy; ///< band energy
+ float thr; ///< energy threshold
+ float thr_quiet; ///< threshold in quiet
float nz_lines; ///< number of non-zero spectral lines
float active_lines; ///< number of active spectral lines
float pe; ///< perceptual entropy
};
/**
- * calculates the attack threshold for ABR from the above table for the LAME psy model
+ * Calculate the ABR attack threshold from the above LAME psymodel table.
*/
static float lame_calc_attack_threshold(int bitrate)
{
/**
* LAME psy model specific initialization
*/
-static void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx) {
+static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx)
+{
int i, j;
for (i = 0; i < avctx->channels; i++) {
AacPsyCoeffs *coeffs = pctx->psy_coef[j];
const uint8_t *band_sizes = ctx->bands[j];
float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
- float avg_chan_bits = chan_bitrate / ctx->avctx->sample_rate * (j ? 128.0f : 1024.0f);
+ float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate;
/* reference encoder uses 2.4% here instead of 60% like the spec says */
float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
* Tell encoder which window types to use.
* @see 3GPP TS26.403 5.4.1 "Blockswitching"
*/
-static FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
- const int16_t *audio, const int16_t *la,
- int channel, int prev_type)
+static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
+ const int16_t *audio,
+ const int16_t *la,
+ int channel, int prev_type)
{
int i, j;
int br = ctx->avctx->bit_rate / ctx->avctx->channels;
AacPsyChannel *pch = &pctx->ch[channel];
uint8_t grouping = 0;
int next_type = pch->next_window_seq;
- FFPsyWindowInfo wi;
+ FFPsyWindowInfo wi = { { 0 } };
- memset(&wi, 0, sizeof(wi));
if (la) {
float s[8], v;
int switch_to_eight = 0;
int stay_short = 0;
for (i = 0; i < 8; i++) {
for (j = 0; j < 128; j++) {
- v = iir_filter(la[(i*128+j)*ctx->avctx->channels], pch->iir_state);
+ v = iir_filter(la[i*128+j], pch->iir_state);
sum += v*v;
}
s[i] = sum;
/**
* Calculate band thresholds as suggested in 3GPP TS26.403
*/
-static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
- const float *coefs, const FFPsyWindowInfo *wi)
+static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
+ const float *coefs, const FFPsyWindowInfo *wi)
{
AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
AacPsyChannel *pch = &pctx->ch[channel];
float desired_bits, desired_pe, delta_pe, reduction, spread_en[128] = {0};
float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
- const int num_bands = ctx->num_bands[wi->num_windows == 8];
- const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
- AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
+ const int num_bands = ctx->num_bands[wi->num_windows == 8];
+ const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
+ AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
//calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
for (w = 0; w < wi->num_windows*16; w += 16) {
for (g = 0; g < num_bands; g++) {
AacPsyBand *band = &pch->band[w+g];
+
float form_factor = 0.0f;
band->energy = 0.0f;
for (i = 0; i < band_sizes[g]; i++) {
band->energy += coefs[start+i] * coefs[start+i];
form_factor += sqrtf(fabs(coefs[start+i]));
}
- band->thr = band->energy * 0.001258925f;
+ band->thr = band->energy * 0.001258925f;
band->nz_lines = form_factor / powf(band->energy / band_sizes[g], 0.25f);
- start += band_sizes[g];
+ start += band_sizes[g];
}
}
//modify thresholds and energies - spread, threshold in quiet, pre-echo control
for (w = 0; w < wi->num_windows*16; w += 16) {
AacPsyBand *bands = &pch->band[w];
- //5.4.2.3 "Spreading" & 5.4.3 "Spreaded Energy Calculation"
+
+ /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */
spread_en[0] = bands[0].energy;
for (g = 1; g < num_bands; g++) {
- bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
+ bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
}
for (g = num_bands - 2; g >= 0; g--) {
- bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
+ bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
}
//5.4.2.4 "Threshold in quiet"
for (g = 0; g < num_bands; g++) {
AacPsyBand *band = &bands[g];
+
band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
//5.4.2.5 "Pre-echo control"
if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
- /* 5.6.1.3.1 "Prepatory steps of the perceptual entropy calculation" */
+ /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */
pe += calc_pe_3gpp(band);
a += band->pe_const;
active_lines += band->active_lines;
}
/* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
- ctx->pe[channel] = pe;
+ ctx->ch[channel].entropy = pe;
desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
/* NOTE: PE correction is kept simple. During initial testing it had very
for (w = 0; w < wi->num_windows*16; w += 16) {
for (g = 0; g < num_bands; g++) {
AacPsyBand *band = &pch->band[w+g];
- FFPsyBand *psy_band = &ctx->psy_bands[channel*PSY_MAX_BANDS+w+g];
+ FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
psy_band->threshold = band->thr;
psy_band->energy = band->energy;
memcpy(pch->prev_band, pch->band, sizeof(pch->band));
}
+static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
+ const float **coeffs, const FFPsyWindowInfo *wi)
+{
+ int ch;
+ FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
+
+ for (ch = 0; ch < group->num_ch; ch++)
+ psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
+}
+
static av_cold void psy_3gpp_end(FFPsyContext *apc)
{
AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
ctx->next_window_seq = blocktype;
}
-static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
- const int16_t *audio, const int16_t *la,
- int channel, int prev_type)
+static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio,
+ const float *la, int channel, int prev_type)
{
AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
AacPsyChannel *pch = &pctx->ch[channel];
int uselongblock = 1;
int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
int i;
- FFPsyWindowInfo wi;
+ FFPsyWindowInfo wi = { { 0 } };
- memset(&wi, 0, sizeof(wi));
if (la) {
float hpfsmpl[AAC_BLOCK_SIZE_LONG];
float const *pf = hpfsmpl;
float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
- int chans = ctx->avctx->channels;
- const int16_t *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN) * chans;
+ const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN);
int j, att_sum = 0;
/* LAME comment: apply high pass filter of fs/4 */
for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
float sum1, sum2;
- sum1 = firbuf[(i + ((PSY_LAME_FIR_LEN - 1) / 2)) * chans];
+ sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2];
sum2 = 0.0;
for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
- sum1 += psy_fir_coeffs[j] * (firbuf[(i + j) * chans] + firbuf[(i + PSY_LAME_FIR_LEN - j) * chans]);
- sum2 += psy_fir_coeffs[j + 1] * (firbuf[(i + j + 1) * chans] + firbuf[(i + PSY_LAME_FIR_LEN - j - 1) * chans]);
+ sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]);
+ sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]);
}
- hpfsmpl[i] = sum1 + sum2;
+ /* NOTE: The LAME psymodel expects its input in the range -32768 to
+ * 32768. Tuning this for normalized floats would be difficult. */
+ hpfsmpl[i] = (sum1 + sum2) * 32768.0f;
}
/* Calculate the energies of each sub-shortblock */
float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
float p = 1.0f;
for (; pf < pfe; pf++)
- if (p < fabsf(*pf))
- p = fabsf(*pf);
+ p = FFMAX(p, fabsf(*pf));
pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
- /* FIXME: The indexes below are [i + 3 - 2] in the LAME source.
- * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
- * (which is what we use here). What the 3 stands for is ambigious, as it is both
- * number of short blocks, and the number of sub-short blocks.
- * It seems that LAME is comparing each sub-block to sub-block + 1 in the
- * previous block.
+ /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
+ * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
+ * (which is what we use here). What the 3 stands for is ambiguous, as it is both
+ * number of short blocks, and the number of sub-short blocks.
+ * It seems that LAME is comparing each sub-block to sub-block + 1 in the
+ * previous block.
*/
if (p > energy_subshort[i + 1])
p = p / energy_subshort[i + 1];
projects / ffmpeg / blobdiff